OpenAI Unveils Dots: Always-On Agents With Their Own Computers — the Agent War Enters Its 'Form Factor' Phase

On September 29, at San Francisco's Fort Mason Center, OpenAI opened DevDay 2026. Sam Altman took the stage and led not with a model, but with a product: Dots — an "always-on" AI assistant. Each Dot gets its own cloud computer and browser, powered by GPT-6 Astra, and works for you around the clock. You give it a goal; it digs through your apps, runs the workflows, and reports back. This isn't another chat box. It's a "digital employee" with its own desk.

The timing is telling. Just three weeks earlier, Meta launched Muse, a free personal agent backed by Facebook, Instagram, and WhatsApp's 3 billion daily active users — overtaking ChatGPT atop the download charts. After four years of dominating the chatbot market, OpenAI suddenly found itself the challenger in the "personal assistant" category. Dots is its counterpunch — and the strategy is completely different: Meta is betting on free and scale; OpenAI is betting on depth and paid subscriptions.

What Dots Actually Is

In OpenAI's words, Dots are "remarkably capable, always-on agents built to handle everything." Three design decisions define the product:

• One "primary dot" per person, with a name and avatar. You can name it and pick its look — the keynote showed a yellow triangle called Alfred and a pink heart called Jojo. For now each account gets one; multiple dots are coming, plus "specialist dots" for organizations.
• It has its own computer. Every Dot runs on a dedicated cloud computer with its own browser, connecting to 4,000+ apps through the plugin ecosystem. You can text or voice-call it from ChatGPT on web, desktop, and mobile, or message it directly in Slack and Microsoft Teams. SMS support is "coming soon."
• It finds work on its own. When you haven't assigned anything, a Dot switches to what OpenAI calls "proactive research": scanning your authorized apps in read-only mode for ways to help. Noticed you forgot to send an invoice? It quietly prepares the draft.

Platformer's Casey Newton, who tried a preview build, wrote that about 15 minutes of setup bought him roughly two hours of work. Altman's pitch on stage was blunter: treat it "the way you would [delegate to] a high agency engineer or a chief of staff" — hand over ambitious chunks of work whole.

Pricing and Access: Lock In the Paying Users First

Dots is launching to paid tiers only: ChatGPT Pro ($100/month) and Business Premium subscribers, with the first Dot included at no extra cost. Enterprise, Education, and Healthcare workspaces can try a beta, but it's off by default until an admin enables it. Pro users in the EEA, Switzerland, and the UK are excluded from this rollout.

One detail matters enormously: conversations with a Dot don't count toward regular ChatGPT usage limits. That's OpenAI drawing a bright line — it wants you to separate "chatting" from "getting things done." Chat keeps its metered pricing; Dots is a separate business line priced on "workload." The keynote also previewed paid upgrades to each Dot's speed or monthly workload. The message is unmistakable: agent compute is about to become OpenAI's next billing dimension.

Safety: Rules Before Actions

Always-on means an agent touching your data and accounts while you sleep — safety is the existential question for this entire product category. OpenAI gave Dots three layers of guardrails: built-in default behavior rules, user-defined custom rules (e.g., "ask me before anything involving payment"), and an auto-review mechanism that checks every planned action against those rules. When a Dot can't safely complete something alone — changing a password, say — it hands the task back to the human.

The design is a direct response to this year's agent security incidents. Coverage of the launch noted that, "reflecting recent security incidents involving agents," Dots ships with custom rule-setting. OpenAI's finance chief added at DevDay that the company will "pace the frontier" when safety requires it — corporate-speak for: we'd rather ship slower than ship another incident.

My Take: The Agent War Enters the "Form Factor" Phase

If 2024 was the war of models and 2025 the war of coding agents, then starting this fall, the war is about form: what shape should AI take in your working life?

Three approaches are now on the table. Meta's Muse: free, personable, armed with the social graph — betting that everyone should have one. OpenAI's Dots: paid, professional, with its own work environment — betting it's worth paying a salary for. And Microsoft's Autopilot (announced last week): living directly inside the corporate directory with an identity, permissions, and audit trails — betting on winning IT departments first. Three form factors, three business models: ads/scale, subscription/workload, enterprise licensing.

For developers and indie builders, Dots has an immediate implication: for the first time, you can hire a 24/7 "junior engineer." Monitoring user feedback, turning issues into code changes, running regression tests, recording demo videos — that's exactly the pipeline OpenAI demoed. One person plus one Dot can theoretically sustain the operate-and-iterate cadence that used to need a 2–3 person team.

But hold on for three seconds, because always-on comes with three bills to pay. First, the security bill: an agent that can read all your apps, drives a browser, and acts on its own at 3 AM — your permission boundaries are your entire net worth. How well you write those custom rules determines whether it's an assistant or an insider threat. Second, the cost bill: "conversations don't count toward usage" sounds generous, but tasks a Dot kicks off in Codex or ChatGPT Work are metered as usual — a cloud computer running 24/7 burns money every minute. Third, the attention bill: when the agent starts finding its own work, you save execution time but gain review time. Newton's "15 minutes for 2 hours" is seductive, but those 15 minutes carry far more decision density than two hours of scrolling.

My conclusion: Dots' biggest contribution isn't "doing your work for you" — it's the first productization of the employment relationship with an agent: naming it, setting goals, defining permissions, watching it work, intervening when needed. Once this interaction paradigm sticks, every coding agent will follow. For vibe coders, the time to practice "being a manager" is now: write crisp goals, define acceptance criteria, learn to review agent output. Because in the coming years, your edge may not be how fast you write code — it'll be how well you run your agent team.

DevDay also packed 20+ other announcements: GPT-6.1 Sol (roughly 20% of Astra's token price), an Ultrafast speed tier, the ChatGPT Space collaboration workspace — and the disclosure that ChatGPT now reaches 1.2 billion weekly users. But a decade from now, what people may remember about this day is simpler: it was the day the AI assistant got a job description, and you became an agent's boss for the first time.